ohwx is a word with no meaning, and that is the entire point of it. It is a rare token: a short string that an image model has almost no prior knowledge of, which makes it the perfect name to give something the model has never seen before.
Why a model needs a made-up name
Suppose you want a text-to-image model to learn your face. You show it twenty photographs and you have to tell it what to call you. If you pick your own name, you inherit every association the model already has with it — every other person of that name in its training data, and whatever the word means in other languages. The model spends its effort fighting what it already believes.
Google Research’s DreamBooth paper set out the solution in 2022. You bind the subject to an identifier and a class noun, in prompts shaped like a photo of [V] man. The class noun — man, dog, chair — tells the model what kind of thing this is, so it can keep everything it already knows about people or dogs. The identifier is the new part, and it works best when the model has never had an opinion about it. The paper’s authors went looking for rare tokens in the vocabulary precisely because ordinary English words come pre-loaded with meaning the model then has to be talked out of.
The problem with sks
The identifier most people met first was sks. It appeared in the example code that early implementations were built from, and it stuck the way defaults do — copied from one tutorial to the next by people who reasonably assumed it had been chosen for a reason.
It had not. SKS is a Soviet semi-automatic carbine, and Stable Diffusion knows it as one. Train a person on sks man and rifles start turning up in the output: in the subject’s hands, on the wall behind them, as a shape the model reaches for because part of the prompt still means gun to it. The identifier that was supposed to be empty turned out to be full.
This is the failure that makes the whole idea legible. A rare token is not a stylistic preference. It is a way of making sure the name you are teaching does not already point somewhere else.
Why ohwx works
Once the sks problem was understood, people went through the tokenizer looking for strings that were genuinely unclaimed — short, with no dictionary meaning, no brand behind them, no strong visual association anywhere in the model. ohwx is one of the ones that came out of that search and stayed, and you will find it across thousands of public fine-tunes, almost always in the same shape: ohwx man, ohwx woman, ohwx person.
We should be straight about one thing: the origin of sks can be traced, but ohwx cannot. It spread through community tooling, forum threads and training guides rather than through any paper or release, and we have not found a single first use worth pointing at. If you know where it started, tell us and we will correct this page.
It did not go away with DreamBooth
Full DreamBooth fine-tuning is no longer how most people personalise a model — LoRA and its relatives are lighter and faster. But the underlying need did not change. A LoRA still has to be triggered by something, and that trigger is still a token the base model ideally knows nothing about. The vocabulary of trigger words in use today descends directly from this problem, and the reasoning is identical: if the name means something already, the meaning leaks.
Why the site is called ohwx.ai
This site is about generative models: what they learned, what they did not, and what happens at the edge between the two. ohwx is the word the field invented for exactly that edge — the placeholder you hand a model for a thing it has not learned yet.
It is also, we will admit, the kind of name that means nothing to almost everyone and something quite specific to the people who have trained a model at three in the morning and wondered why their portraits kept coming out holding a rifle.
